Identifying PTSD sex-based patterns through explainable artificial intelligence in biometric data

Abstract

Post-Traumatic Stress Disorder (PTSD) is a mental health condition that arises from exposure to traumatic events, affecting various aspects of human well-being. The complexity and variability of symptoms pose challenges for accurate diagnosis and monitoring, exacerbated by accessibility barriers. In response, alternative methodologies leveraging biometric data have emerged, such as facial movements, speech, or voice-to-text transcriptions, and analyzing them using Explainable Artificial Intelligence (XAI) techniques. Numerous studies have explored the presence or absence of PTSD, yet few have concentrated on either explicable indicators or distinctions among these indicators, such as the patient’s sex. This research used an XAI algorithm to identify patterns related to PTSD in three biometric data sets: facial movements, speech, and voice-to-text transcriptions. Such biometric data are grouped by sex. Utilizing the DAIC-WOZ database, this study involves feature selection and characterization. Experiment configurations addressed participant segmentation, feature reduction, and standardization. Training phases employed machine-learning classification algorithms with their corresponding performance evaluation. The interpretability stage explored the relationship between input features and class output. The findings reveal that among the three biometric data sets evaluated in this work (facial movements, speech, and voice-to-text transcriptions), speech characterization is the most effective in identifying PTSD indicators, suggesting a uniform speech pattern associated with breathy and tense voice and weak phonation in PTSD patients. Sex-specific analysis enhances prediction performance, revealing distinctions in the associated speech features. The Women’s model prioritizes tense voice and vocal volume variations, reduced glottal closure, and interrupted phonation. Conversely, the Speech-Men model reflects reduced resonance, making the voice thinner and weaker, indicating altered vocal quality. As for facial movements, sex-specific characteristics are not evident, but some features focused on lips are associated with PTSD. Similarly, PTSD is related to alertness, determination, and anxiety in both women and men. In conclusion, using an XAI algorithm to differentiate sex-based patterns in biometric data contributes to a better understanding of PTSD indicators, offering potential advancements in personalized diagnostic strategies.